交通工程

为提升不安全驾驶行为预判能力，本研究基于多模态生理与驾驶行为数据，提出融合特征筛选与多模型对比的预测方法。首先，利用LightGBM结合5折交叉验证进行动态特征筛选，从30维原始特征中提取出8个核心人因与行为特征变量；其次，构建包含ANN、DT、ET、KNN和XGBoost等9种模型的对比体系，通过网格搜索优化参数，并基于ROC曲线、AUC、准确度、敏感度、特异度等指标评估性能；进一步集成SHAP与PCA技术揭示特征作用机制。实验表明，ANN模型综合性能最优，测试集准确率达92.8%,F1-score为0.929,AUC值达0.979，显著优于传统集成方法。SHAP解释结合PCA可视化揭示了眼球注视率、脑电信号等关键人因特征对不安全行为的非线性影响，并在特征空间展现出显著类别可分性。研究证实了多模型对比分析在平衡预测性能与可解释性方面的有效性，为智能驾驶系统人机交互及行为预判模块优化提供了模型选择依据。

To enhance the predictive capability for unsafe driving behaviors, this study proposes a prediction method integrating feature selection and multi-model comparison based on multimodal physiological and driving behavior data. Firstly, LightGBM combined with 5-fold cross-validation was employed for dynamic feature selection, extracting 8 core human factor and behavioral characteristic variables from the original 30-dimensional features. Secondly, a comparative framework comprising 9 models, including ANN, DT, ET, KNN, and XGBoost, was constructed. Model parameters were optimized via grid search, and performance was evaluated using metrics such as the ROC curve, AUC, accuracy, sensitivity, and specificity. Furthermore, SHAP and PCA techniques were integrated to elucidate the underlying mechanisms of the features. Experiments demonstrated that the ANN model exhibited the best comprehensive performance, achieving a test set accuracy of 92.8%, an F1-score of 0.929, and an AUC value of 0.979, significantly outperforming traditional ensemble methods. SHAP interpretation combined with PCA visualization revealed the nonlinear effects of key human factor features, such as eye gaze fixation rate and EEG signals, on unsafe behaviors, and demonstrated significant class separability in the feature space. This study confirms the effectiveness of multi-model comparative analysis in balancing predictive performance and interpretability, providing a basis for model selection in optimizing human-computer interaction and behavior prediction modules within intelligent driving systems.